We test whether these limitations can be explained through two key failures of the self-attention mechanism. The first is the inability to fully remove irrelevant information. The second is tied to position, even if the dot product between a key and query is highly negative (i.e. an irrelevant key) learned positional biases may unintentionally up-weight such information - dangerous when distances become out of distribution. Put together, these two failure cases lead to compounding generalisation difficulties. We test whether they can be mitigated through the combination of a) selective sparsity - completely removing irrelevant keys from the attention softmax and b) contextualised relative distance - distance is only considered as between the query and the keys that matter. We show how refactoring the attention mechanism with these two mitigations in place can substantially improve the generalisation capabilities of decoder only transformers.

Diffusion Probabilistic Models (DPMs) have been recently utilized to deal with various blind image restoration (IR) tasks, where they have demonstrated outstanding performance in terms of perceptual quality. However, the task-specific nature of existing solutions and the excessive computational costs related to their training, make such models impractical and challenging to use for different IR tasks than those that were initially trained for. This hinders their wider adoption, especially by those who lack access to powerful computational resources and vast amount of training data. In this work we aim to address the above issues and enable the successful adoption of DPMs in practical IR-related applications. Towards this goal, we propose a modular diffusion probabilistic IR framework (DP-IR), which allows us to combine the performance benefits of existing pre-trained state-of-the-art IR networks and generative DPMs, while it requires only the additional training of a relatively small module (0.7M params) related to the particular IR task of interest. Moreover, the architecture of the proposed framework allows for a sampling strategy that leads to at least four times reduction of neural function evaluations without suffering any performance loss, while it can also be combined with existing acceleration techniques such as DDIM. We evaluate our model on four benchmarks for the tasks of burst JDD-SR, dynamic scene deblurring, and super-resolution. Our method outperforms existing approaches in terms of perceptual quality while it retains a competitive performance with respect to fidelity metrics.

While language models (LMs) paired with residual vector quantization (RVQ) tokenizers have shown promise in text-to-audio (T2A) generation, they still lag behind diffusion-based models by a non-trivial margin. We identify a critical dilemma underpinning this gap: incorporating more RVQ layers improves audio reconstruction fidelity but exceeds the generation capacity of conventional LMs. To address this, we first analyze RVQ dynamics and uncover two key limitations: 1) orthogonality of features across RVQ layers hinders effective LMs training, and 2) descending semantic richness in tokens from deeper RVQ layers exacerbates exposure bias during autoregressive decoding. Based on these insights, we propose Siren, a novel LM-based framework that employs multiple isolated transformers with causal conditioning and anti-causal alignment via reinforcement learning. Extensive experiments demonstrate that Siren outperforms both existing LM-based and diffusion-based T2A systems, achieving state-of-the-art results. By bridging the representational strengths of LMs with the fidelity demands of audio synthesis, our approach repositions LMs as competitive contenders against diffusion models in T2A tasks. Moreover, by aligning audio representations with linguistic structures, Siren facilitates a promising pathway toward unified multi-modal generation frameworks.

Large-scale outbreaks of epidemics, misinformation, or other harmful contagions pose significant threats to human society, yet the fundamental question of whether an emerging outbreak will escalate into a major epidemic or naturally die out remains largely unaddressed. This problem is challenging, partially due to inadequate data during the early stages of outbreaks and also because established models focus on average behaviors of large epidemics rather than the stochastic nature of small transmission chains. Here, we introduce the first systematic framework for forecasting whether initial transmission events will amplify into major outbreaks or fade into extinction during early stages, when intervention strategies can still be effectively implemented. Using extensive data from stochastic spreading models, we developed a deep learning framework that predicts early-stage spreading outcomes in real-time. Validation across Erdős-Rényi and Barabási-Albert networks with varying infectivity levels shows our method accurately forecasts stochastic spreading events well before potential outbreaks, demonstrating robust performance across different network structures and infectivity scenarios.To address the challenge of sparse data during early outbreak stages, we further propose a pretrain-finetune framework that leverages diverse simulation data for pretraining and adapts to specific scenarios through targeted fine-tuning. The pretrain-finetune framework consistently outperforms baseline models, achieving superior performance even when trained on limited scenario-specific data. To our knowledge, this work presents the first framework for predicting stochastic take-off versus die-out. This framework provides valuable insights for epidemic preparedness and public health decision-making, enabling more informed early intervention strategies.

Large Language Models (LLMs) reasoning abilities are increasingly being applied to classical board and card games, but the dominant approach -- involving prompting for direct move generation -- has significant drawbacks. It relies on the model's implicit fragile pattern-matching capabilities, leading to frequent illegal moves and strategically shallow play. Here we introduce an alternative approach: We use the LLM to translate natural language rules and game trajectories into a formal, executable world model represented as Python code. This generated model -- comprising functions for state transition, legal move enumeration, and termination checks -- serves as a verifiable simulation engine for high-performance planning algorithms like Monte Carlo tree search (MCTS). In addition, we prompt the LLM to generate heuristic value functions (to make MCTS more efficient), and inference functions (to estimate hidden states in imperfect information games). Our method offers three distinct advantages compared to directly using the LLM as a policy: (1) Verifiability: The generated CWM serves as a formal specification of the game's rules, allowing planners to algorithmically enumerate valid actions and avoid illegal moves, contingent on the correctness of the synthesized model; (2) Strategic Depth: We combine LLM semantic understanding with the deep search power of classical planners; and (3) Generalization: We direct the LLM to focus on the meta-task of data-to-code translation, enabling it to adapt to new games more easily. We evaluate our agent on 10 different games, of which 4 are novel and created for this paper. 5 of the games are fully observed (perfect information), and 5 are partially observed (imperfect information). We find that our method outperforms or matches Gemini 2.5 Pro in 9 out of the 10 considered games.

Abstract--This paper proposes "3Dify," a procedural 3D computer graphics (3D-CG) generation framework utilizing Large Language Models (LLMs). The framework enables users to generate 3D-CG content solely through natural language instructions. For 3D-CG generation support, 3Dify automates the operation of various Digital Content Creation (DCC) tools via MCP . When DCC tools do not support MCP-based interaction, the framework employs the Computer-Using Agent (CUA) method to automate Graphical User Interface (GUI) operations. Moreover, to enhance image generation quality, 3Dify allows users to provide feedback by selecting preferred images from multiple candidates. The LLM then learns variable patterns from these selections and applies them to subsequent generations. Furthermore, 3Dify supports the integration of locally deployed LLMs, enabling users to utilize custom-developed models and to reduce both time and monetary costs associated with external API calls by leveraging their own computational resources. Its applications extend beyond entertainment industries such as movies and games to areas including product design in manufacturing, surgical simulation in healthcare, education, and digital-twin technologies that replicate the real world within virtual spaces.

Large language models (LLMs) have become integral to enterprise operations, powering applications ranging from automated financial auditing and risk assessment in banking to predictive diagnostics and patient interaction systems in healthcare, and even real-time customer sentiment analysis in e-commerce platforms. However, the deployment of these models at scale introduces multifaceted vulnerabilities that can lead to catastrophic failures. Prompt injection attacks, where malicious inputs manipulate model behavior to bypass safeguards, represent a direct security threat. Strategic deception, where models exhibit emergent behaviors that misalign with intended goals, erodes trust in agentic systems. Biased outputs, stemming from skewed training data or architectural inductive biases, perpetuate unfairness and can result in regulatory non-compliance or reputational damage. Our prior work [Ravindran, 2024] laid the groundwork by introducing adversarial activation patching, a novel interpretability technique that successfully induced deception in simplified toy neural networks, achieving a 23.9% induction rate. This demonstrated the feasibility of using activation-level interventions to probe and expose hidden risks in safety-aligned transformers. Building upon this foundation, we propose the Unified Threat Detection and Mitigation Framework (UTDMF), a comprehensive, scalable, and real-time pipeline explicitly designed for enterprise environments where high-stakes decisions demand robustness, explainability, and compliance.

Accurate auto-formalization of theorem statements is essential for advancing automated discovery and verification of research-level mathematics, yet remains a major bottleneck for LLMs due to hallucinations, semantic mismatches, and their inability to synthesize new definitions. To tackle these issues, we present Aria (Agent for Retrieval and Iterative Autoformalization), a system for conjecture-level formalization in Lean that emulates human expert reasoning via a two-phase Graph-of-Thought process: recursively decomposing statements into a dependency graph and then constructing formalizations from grounded concepts. To ensure semantic correctness, we introduce AriaScorer, a checker that retrieves definitions from Mathlib for term-level grounding, enabling rigorous and reliable verification. We evaluate Aria on diverse benchmarks. On ProofNet, it achieves 91.6% compilation success rate and 68.5% final accuracy, surpassing previous methods. On FA TE-X, a suite of challenging algebra problems from research literature, it outperforms the best baseline with 44.0% vs. 24.0% On a dataset of homological conjectures, Aria reaches 42.9% final accuracy while all other models score 0%. In recent years, Interactive Theorem Provers (ITPs) such as Coq (Barras et al., 1999), Isabelle (Paul-son, 1994) and Lean (Moura & Ullrich, 2021) have become crucial ecosystems for formalized mathematics. Among these, Lean 4, together with its comprehensive library Mathlib (mathlib Community, 2020), is pioneering a new paradigm for formalization. However, the continuous growth of this ecosystem is always constrained by the immense manual effort and the deep expertise that formalization demands. To address this, the research community has turned to Large Language Models (LLMs) for auto-formalization the process of translating informal (or natural language) mathematical statements and proofs into their formal counterparts. While these two processes are interconnected, the accurate formalization of statements is the foundational first step. A correctly formalized statement is a prerequisite for any valid proof and, on its own, is a valuable asset to the mathematical ecosystem, enabling better search, integration, and verification. Thus, despite progress in proof automation (Ren et al., 2025; Chen et al., 2025), the fidelity of this initial statement translation remains a critical bottleneck. LLMs frequently generate formal statements that suffer not only from compilation errors but also from more insidious semantic flaws, a challenge that intensifies when formalizing more complex research or conjecture-level statements.

Prevailing methods for training Large Language Models (LLMs) as text encoders rely on contrastive losses that treat the model as a black box function, discarding its generative and reasoning capabilities in favor of static embeddings. We introduce GRACE (Generative Representation Learning via Contrastive Policy Optimization), a novel framework that reimagines contrastive signals not as losses to be minimized, but as rewards that guide a generative policy. In GRACE, the LLM acts as a policy that produces explicit, human-interpretable rationales--structured natural language explanations of its semantic understanding. These rationales are then encoded into high-quality embeddings via mean pooling. Using policy gradient optimization, we train the model with a multi-component reward function that maximizes similarity between query positive pairs and minimizes similarity with negatives. This transforms the LLM from an opaque encoder into an interpretable agent whose reasoning process is transparent and inspectable. On MTEB benchmark, GRACE yields broad cross category gains: averaged over four backbones, the supervised setting improves overall score by 11.5% over base models, and the unsupervised variant adds 6.9%, while preserving general capabilities. This work treats contrastive objectives as rewards over rationales, unifying representation learning with generation to produce stronger embeddings and transparent rationales. The model, data and code are available at https://github.com/GasolSun36/GRACE.

GenQuest is a generative text adventure game that leverages Large Language Models (LLMs) to facilitate second language learning through immersive, interactive storytelling. The system engages English as a Foreign Language (EFL) learners in a collaborative "choose-your-own-adventure" style narrative, dynamically generated in response to learner choices. Game mechanics such as branching decision points and story milestones are incorporated to maintain narrative coherence while allowing learner-driven plot development. Key pedagogical features include content generation tailored to each learner's proficiency level, and a vocabulary assistant that provides in-context explanations of learner-queried text strings, ranging from words and phrases to sentences. Findings from a pilot study with university EFL students in China indicate promising vocabulary gains and positive user perceptions. Also discussed are suggestions from participants regarding the narrative length and quality, and the request for multi-modal content such as illustrations.